Combining users' activity survey and simulators to evaluate human activity recognition systems.

Area of Science:

• Computer Science
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Evaluating human activity recognition (HAR) systems typically involves resource-intensive human subject experiments.
• These traditional methods raise ethical and legal concerns, alongside significant time and cost investments.

Purpose of the Study:

• To propose a novel, efficient, and ethical evaluation methodology for human activity recognition systems.
• To reduce the reliance on expensive and time-consuming experiments with human participants.

Main Methods:

• The proposed methodology utilizes user surveys to gather data on daily living activities.
• A synthetic dataset generator tool creates labeled activity datasets based on survey information.
• The tool simulates sensor noise, varying time lapses, and erratic user behavior for realistic data.

Main Results:

• A synthetic dataset generated via the new methodology was compared to a real-world dataset.
• Similarity was calculated by comparing sensor occurrence frequencies between the synthetic and real datasets.
• A highly significant similarity was found between the generated synthetic data and the real-world data.

Conclusions:

• The novel evaluation methodology offers significant advantages for researchers, enabling more efficient work.
• The synthetic dataset generator effectively models real-world activity data, validating the proposed approach.
• This method provides a viable alternative to traditional human-subject-based evaluations for HAR systems.